Trust needed evidence, not reassurance

01 / THE ADOPTION RISK

An expert has to be able to defend the result

BuildTwin was adding AI quality control to structural drawings. A quality engineer would review the results and raise issues before an approver moved the work forward.

AI could perform checks, but accountability stayed with the engineer. If the result could not be inspected and corrected, the older manual process remained easier to defend.

The brief arrived as a compact checklist inside an existing drawing viewer. The deeper problem was whether an expert could trust what appeared inside it.

02 / THE RESET

Back from screens to the work

Three roles were mapped: the detailer who created the drawing, the quality engineer who reviewed it, and the approver who moved it forward.

The AI journey introduced one decisive stage. The quality engineer had to review the system’s work before acting on it. That stage became the centre of the design.

03 / THE TRUST PROBLEM

Confidence cannot be declared

A reviewer needs enough evidence to decide whether confidence is deserved.

Each result exposed the drawing reference, source document, observation and conclusion. If information was missing, the interface showed what the reviewer needed to add before running the check again.

“We need to train the user to trust the system, but in the beginning, they won’t trust it.”

BUILDTWIN STAKEHOLDER

04 / THE DECISION

A checklist that behaves like a feed

Accordions kept the panel compact, but hid evidence behind repeated clicks. A continuous feed used more space and kept every observation, source and status in the reviewer’s path.

The selected direction made complete review the default. Search and filters remained available, but evidence no longer disappeared when the reviewer moved between checks.

“I don’t want a user to, let’s say, I don’t like this clicking.”

BUILDTWIN STAKEHOLDER

05 / HUMAN CONTROL

Authority belongs at the point of decision

The reviewer could add missing inputs, inspect reasoning, override a status and raise an issue against the observation that produced it.

An override required a corrected state and a reason. Human judgement remained visible rather than silently replacing the model’s answer.

06 / WHERE IT STANDS

A product model, not a shipped outcome

Across repeated reviews, the model covered checks in progress, missing inputs, evidence, reasoning, overrides, issues and history.

The available material does not establish what subsequently shipped. Adoption, time saved and accuracy improvement are not claimed here.

WHAT THIS PROVES

AI products in consequential domains need to show their work and leave control where accountability already sits.

The expensive mistake would have been finishing the first interface faster without resolving what the expert needed to trust, correct and escalate.